TL;DR
This paper introduces a novel method to detect if a High-Efficiency Video Coding (HEVC) video has been previously transcoded from an Advanced Video Coding (AVC) format. By analyzing the statistical patterns of Prediction Units (PUs) within HEVC frames, the method achieves high accuracy in distinguishing original HEVC videos from those that have undergone a transcoding process, offering a valuable tool for video forensics.
What Problem Does This Paper Address?
The increasing prevalence of digital video, especially on online platforms, means videos are frequently edited, compressed, and re-encoded. This process can obscure the video's original compression history, making it difficult to determine if a video is authentic or has been manipulated. Specifically, older or lower-quality AVC videos can be re-encoded into the newer, more efficient HEVC format, presenting them as high-quality originals. This misrepresentation can be problematic for various applications, including digital forensics, content authentication, and quality assessment. While methods exist for detecting transcoded videos in older formats like MPEG, there's a notable gap in research for the latest HEVC standard, particularly when the source is AVC. This paper aims to fill this gap by developing a technique to identify HEVC videos that have been transcoded from AVC.
Methods
The core of the proposed method lies in analyzing the "footprints" left by the transcoding process on the HEVC video's structure. The authors observed that the AVC encoding process influences the PU partitioning strategy in the subsequent HEVC encoding. Specifically, when an AVC video is transcoded to HEVC, the resulting PU partitions in the HEVC stream differ statistically from those in a natively encoded HEVC video.
The method involves the following steps:
- Decoding and Parsing: The HEVC video is decoded, and its bitstream is parsed to extract PU partition data for each frame.
- Feature Extraction:
- I-frames: For intra-coded frames (I-frames), where prediction is solely within the frame, a 5-dimensional feature set ($F_I$) is extracted based on the frequencies of five specific PU sizes.
- P-frames: For inter-coded frames (P-frames), which utilize temporal prediction, a 25-dimensional feature set ($F_P$) is extracted based on the frequencies of twenty-five different PU sizes.
- Feature Combination: The 5D and 25D feature sets are combined to form a comprehensive 30-dimensional feature vector ($F$).
- Classification: A pre-trained Support Vector Machine (SVM) classifier is used to distinguish between original and transcoded HEVC videos based on the 30D feature vector.
The key novelty lies in exploiting the statistical differences in PU partitioning caused by the AVC-to-HEVC transcoding, rather than relying on general compression artifacts.
Datasets
The experiments were conducted on a dataset comprising:
- Originally-encoded HEVC videos: These serve as negative instances.
- Transcoded HEVC videos (from AVC): These serve as positive instances.
The dataset was generated using uncompressed YUV sequences. For CIF resolution videos, short clips of 100 frames were used, resulting in 105 CIF clips from 18 YUV sequences. For High Definition (HD) 720p videos, 75 clips were generated.
The encoding was performed using:
- HEVC: HM 12.0 (High Efficiency Video Coding Test Model) with a low-delay coding mode.
- AVC: JM 19.0 (H.264/AVC Reference Software) with the main profile.
To create transcoded videos, YUV clips were first encoded to AVC using JM, then decoded, and finally re-encoded to HEVC using HM. For negative instances, YUV clips were directly encoded to HEVC using HM.
Four different bitrates were used for encoding: 100 Kbps, 200 Kbps, 300 Kbps, and 400 Kbps. For HD videos, higher bitrates were used: 2 Mbps, 3 Mbps, 4 Mbps, and 5 Mbps.
Key Insights and Ideas
- PU Partitioning as a Forensic Footprint: The study reveals that the choice of PU partitioning in HEVC is sensitive to its compression history. Transcoding from AVC leaves a distinct statistical signature in PU frequencies that can be exploited for detection.
- I-frame vs. P-frame Features: Features extracted from I-frames ($F_I$) are more robust to low-bitrate compression, while features from P-frames ($F_P$) contribute more significantly at higher bitrates. This highlights the complementary nature of these feature sets.
- Combined Features are Superior: Combining features from both I-frames and P-frames ($F$) leads to superior detection accuracy compared to using either set alone, demonstrating the benefit of a holistic approach.
- Robustness to Encoding Parameters: The proposed method shows good robustness to variations in encoding parameters like frame structure and intra-frame distance, suggesting its practical applicability.
Key Results
The proposed method achieved high detection accuracies:
- CIF Videos: The combined feature set ($F$) achieved an average detection accuracy of 98.61% across various bitrate combinations. In some cases, accuracies reached 100.00%.
- HD 720p Videos: Detection accuracies were consistently high, with an average detection rate of 98.17%. The accuracy increased with higher HEVC bitrates, reaching up to 98.85%.
- Comparison with Yu et al. [19]: The proposed method outperformed a recent method by Yu et al. by an average of 5.61% on CIF videos, demonstrating its improved effectiveness.
- Robustness: The method maintained an average detection rate of 91.40% even with altered encoding parameters, showing a decrease of only 2.8% compared to the baseline. Against saturation enhancement, the average detection rate was 93.96%.
Limitations and Future Work
Limitations:
- The current method focuses on AVC-to-HEVC transcoding. Its performance on other transcoding scenarios (e.g., HEVC-to-AVC, or transcoding between different versions of AVC or HEVC) is not discussed.
- The method's performance might be affected by significant video enhancement operations applied before transcoding, although it showed robustness to saturation enhancement.
Future Work:
- Broader Coding Standards: Extend the method to detect transcoding from other previous coding standards like MPEG-2 and MPEG-4.
- Other Encoding Tools: Validate the method using different HEVC encoders (e.g., x265) to assess its generalizability.
- New Features: Explore additional features, such as quantization coefficient prediction errors, to further enhance detection accuracy and adaptability.
Summary updated